Elon Musk's latest interview: Starlink phones will be launched in two years, and Mars colonization will be achieved within 25 years.

According to a report by Zhidx on September 10th, this morning, the latest interview video of Elon Musk at the All-In Technology Summit was released. In this interview with an extremely high information density, Musk detailed his development plans for multiple companies such as Tesla, SpaceX, xAI, and Starlink, covering a wide range of topics including humanoid robots, Starlink phones, Grok, AI chips, super data centers, Martian colonization, and fertility rates. It might be one of Musk's most comprehensive and in - depth public talks recently.

At the beginning of the conversation, Musk sharply criticized his "side mission" of working at the U.S. Government Efficiency Department (DOGE), saying that the U.S. government is almost "hopeless," and that AI and robots might be the hope to save the situation.

Musk revealed that the Optimus 3 humanoid robot will have hand dexterity close to that of humans, with an estimated minimum cost of about $20,000 per unit (equivalent to about 142,400 RMB). He emphasized that this will be a major product that will influence human history. When talking about the difficulty of robot R & D, he believed that the real challenge lies in hardware design, especially the complex engineering problems of the hands and forearms.

When it comes to Tesla, Musk said that the next - generation in - vehicle inference chip, AI5, will achieve a 40 - fold performance leap compared to AI4. The software upgrade at the end of this year is expected to make Tesla cars show "consciousness" without hardware upgrades.

This month, Musk spent a whopping $17 billion to acquire some communication frequency bands of the telecommunications operator Echo Star. This move paves the way for the "Starlink phone," which can directly connect to Starlink without relying on a Starlink router. Musk said that the Starlink phone is expected to be launched in two years, ultimately achieving high - bandwidth connectivity globally. To achieve this goal, he believes that acquiring Verizon, the largest telecommunications operator in the United States, is also under consideration.

As for SpaceX, Musk said that the carrying capacity of the third - generation Starship will exceed 100 tons, and it is expected to be fully reusable next year. Currently, creating a stable and reusable heat shield is the biggest challenge to achieving the above goal.

In the field of AI content generation, he revealed that Grok is rewriting Wikipedia to eliminate false information and supplement missing content. After the host's suggestion, Musk also said that he would consider creating a public "Grok Encyclopedia."

The entire interview ended with the grand topic of human civilization and interstellar exploration. Musk believes that within 25 years at the earliest, humans are expected to achieve self - sufficient colonization on Mars without relying on supplies from Earth. He called on humans to remain curious about the universe, which will open up a brand - new understanding of the world.

Here is the complete translation of Musk's latest interview (Zhidx made some additions, deletions, and modifications to some content without changing the original meaning):

01. Optimus 3 Will Have Human - Level Dexterous Hands, with a Minimum Cost of $20,000 per Unit

Host: Where are you now?

Musk: At Tesla's global engineering headquarters in Palo Alto, California, USA.

Host: You're no longer in Washington and have returned to your main job?

Musk: Yes, I haven't been to Washington since May.

Host: That was quite a remarkable "side mission." What did you learn from your experience in Washington?

Musk: The U.S. government is basically hopeless. There are talented people in the government, which is of course good. But if you look at the high national debt of the United States, the interest payments even exceed the budget of the Department of Defense (or the "War Department"), and it's still rising.

In this situation, if AI and robots can't solve the national debt problem, we're doomed.

Host: I think the Optimus humanoid robot will become the greatest product in human history. How is it progressing now? How much of your energy is spent on Optimus? What's the product timeline? I remember it's already the third - generation, or even the fourth - generation, right?

Musk: It would take a long time to explain the whole situation. We're perfecting the design of Optimus 3, which will be a truly remarkable robot. It will have hand dexterity comparable to that of humans, that is, very complex hands.

It will have an AI brain that can understand and respond to reality, and it will be mass - produced. These three aspects are currently lacking in other robot companies and are the most difficult.

I'm currently spending more mental energy on Optimus than on any other single project. We need to solve AI technology problems, electromechanical technology problems, and supply - chain and production challenges.

Currently, there is no supply chain for humanoid robots at all. We have to rebuild it from scratch, which requires a large amount of vertical integration. All the actuators of Optimus are designed and manufactured by ourselves and can't be bought from the existing supply chain.

If Optimus succeeds, it will become the most influential product in history.

Host: What about the cost? Is it $20,000, $30,000, or $40,000 per unit? How much will the first batch cost?

Musk: When the annual production reaches one million units, the marginal production cost will be around $20,000. It specifically depends on how much is spent on AI chips and the efficiency of the actuators.

Currently, each arm has 26 actuators (including motors, gearboxes, and electronic controls, etc.), and the AI chip itself may cost $5,000 to $6,000 or even more. I estimate that the cost during mass production will be about $20,000 to $25,000, and the selling price will depend on market demand.

02. Building Robots Is More Difficult Than Building Cars, and the Main Difficulty Lies in Hardware Design

Host: Can you explain to everyone why the "hands" are so important, and why the actuator design is so unique and difficult? Why must we start with the hands to build a qualified robot?

Musk: The human hand is an extremely exquisite evolutionary result. The human hand can do many things: clapping, playing baseball, threading a needle, playing the piano, playing the violin, assembling or disassembling a car. The hand is an extremely versatile tool.

Most of the muscles that drive the human hand are actually in the forearm. The human hand is a bit like a marionette, with the muscles in the forearm pulling the ligaments to control the hand. The human hand has about 27 or 28 degrees of freedom. To make a robot a truly general - purpose humanoid robot, the "hand" problem must be solved.

Host: Is this like when you built Tesla, where there was no supply chain at all and you had to build everything yourself?

Musk: Yes, we simply can't buy the required actuators, no matter how much we're willing to pay. Although there are thousands of types of motors on the market, we have to start from the first principles of physics and design the motors, gearboxes, and control electronics from scratch.

Host: Compared with building the Cybertruck, Model Y, Superfactory, or Model X, how difficult do you think it is?

Musk: More difficult. It's more difficult than all of them.

Host: Is it more difficult than the Starship?

Musk: The Starship is more difficult.

Host: Optimus is between the Model X and the Starship. Which is more difficult, hardware or software?

Musk: Currently, the main difficulty lies in hardware design. Especially the hands and forearms. It's not that the other parts aren't difficult, but the engineering difficulty of the hands and forearms is the greatest.

Host: Assuming the hardware problem is solved, can the software directly leverage the progress of large models? For example, allowing consumers to talk to the robot and give commands, and the robot can accurately understand them.

Musk: Yes, no problem.

Host: Why choose the humanoid form? You could build machines that are perhaps better than humans, or simpler than humans, can complete specific tasks, or even do more things than humans.

Musk: If you want a robot to do everything humans can do, you need a humanoid robot. If you only want it to do some of those things, it's much easier.

However, there are good reasons behind human evolution to our current form and abilities. Having four fingers and a thumb is very valuable, and even the little finger is useful. But it's harder to say about the toes.

Host: Humans have also designed the world. If we build a humanoid robot, it will be compatible with the world we've created.

03. AI5 Will Achieve a 40 - Fold Performance Leap, and Tesla Cars Will Have "Consciousness" by the End of the Year

Host: Robots need AI models, actuators, dexterous hands, etc., but they also need chips. You mentioned Dojo, AI5, and AI6 in a tweet, and you seem very excited about the progress. Can you tell us what they are?

Musk: At Tesla, we have two different chip projects. One is Dojo for training, and the other is the inference chip we call AI4. AI4 is currently being shipped to all models, and we're perfecting the design of AI5. Compared with AI4, AI5 will make a huge leap.

In some metrics, AI5 will be 40 times better than AI4, thanks to the close cooperation between the AI software team and the AI hardware team. We know where the limitations are, and the two teams are jointly designing the chip.

Host: Where will this performance improvement be reflected when it's put into production and actually implemented?

Musk: To be precise, the 40 - fold improvement is for the weak points of AI4, that is, running the softmax operation. Currently, AI4 has to go through about 40 steps to run the softmax in simulation mode. In AI5, such operations can be completed natively in just a few steps. The AI5 chip can also dynamically process mixed - precision models.

In terms of nominal raw computing power, AI5 has 8 times the computing power of AI4, about 9 times the memory, and about 5 times the memory bandwidth. But because we're solving some core limitations in AI4, multiplying the 8 - fold increase in computing power by a 5 - fold improvement is where the 40 - fold improvement comes from.

That being said, I'm confident that the current autopilot in Tesla cars powered by the AI4 chip is at least 2 to 3 times safer than human driving, and perhaps even up to 10 times safer. The software to be released in the next few months is designed for this purpose.

Version 14 will be the biggest upgrade to Tesla's autopilot system since Version 12. We'll increase the number of model parameters by an order of magnitude and introduce a large amount of reinforcement learning.

You can think of AI as a compression of reality, but some of the compression steps cause information loss. We've solved these loss problems, and all the improvements will be achieved through software updates. By the end of this year, your Tesla car will show a sense of "consciousness."

04. Starlink Phone Expected to Launch in Two Years, Acquisition of Verizon Not Ruled Out

Host: I heard you spent about $17 billion on buying communication frequency bands. That's not a small amount. This deal is to enable satellites and the Starlink network to directly connect with mobile phones. What will it be like in one or two years? Can we cancel our Verizon accounts and only use Starlink accounts?

We hope so because Verizon's service is really not good. How many people here want a Starlink phone? Everyone raised their hands.

Musk: This is a long - term thing. This deal allows SpaceX to directly provide high - bandwidth connectivity from satellites to mobile phones, but the phones need hardware adjustments. Since current phones don't support these frequency bands, the phone's chipset needs to be modified to add support for these frequencies. I think this kind of phone may be gradually launched in two years.

Meanwhile, we also need to launch satellites that can communicate on these frequency bands. We're manufacturing satellites and collaborating with mobile phone manufacturers to integrate these frequency bands into mobile phones. After completion, satellites and mobile phones will be able to communicate smoothly, achieving high - bandwidth connectivity.

Ultimately, you'll be able to watch high - definition videos on your phone at any time. In indoor environments, these frequency bands can work normally in ordinary houses, but they may be affected in buildings with thick metal roofs.

Host: Is your vision to make us no longer rely on regional operators like AT&T and not need roaming, but directly use Starlink for global connectivity? Is the ultimate goal for Starlink to become a global operator?

Musk: That's one option. We won't make other operators disappear, as they still have a large number of frequency bands. But you'll be able to have a Starlink account, just like using AT&T, T - Mobile, or Verizon. This account can be used for the Starlink antenna at home to provide Wi - Fi and also for mobile phones to achieve comprehensive high - bandwidth connectivity.

Host: As for buying more frequency bands, is it possible to acquire an operator like Verizon in the future?

Musk: It's not impossible. I guess it might happen.

05. Third - Generation Starship Has a Carrying Capacity of Over 100 Tons and Can Be Fully Reused

Host: Let's talk about the Starship. You just completed a very successful launch. How far is the Starship from being predictable and usable in a commercial environment?

Musk: We plan to recover the spacecraft next year. We have one more launch of the second - generation Starship left, and then we'll launch the third - generation Starship. This is a major upgrade because the third - generation Starship is equipped with Raptor 3 engines, and almost all other parts have been redesigned.

Due to the radical design, the third - generation Starship may have some teething problems in the initial stage. But it's fully reusable and can carry over 100 tons of cargo into orbit. Unless there's a major setback, SpaceX will be able to achieve full reusability next year: it can catch both the booster and the spacecraft and carry over 100 tons of cargo into a useful orbit.

Host: What's the carrying capacity of the best rocket in the world currently?

Musk: For commercial rockets, the Falcon Heavy can carry about 40 tons with reusable side boosters. In comparison, the carrying capacity of the Starship is about 2.5 times that.

Host: After the recent launch failure and explosion, there were many questions about the environment, the Federal Aviation Administration, and regulations. But you quickly resumed the launch work, not only recovering technically but also getting the regulatory approval quickly. How did you do that?

Musk: There were indeed many problems to solve. Our speed is mainly due to the SpaceX team. I'm proud of them for being able to recover so quickly. Building a fully reusable orbital rocket is one of the most complex engineering challenges to date and one of the most difficult engineering projects in history.

This has been SpaceX's goal since 2002, and 23 years have passed. It's been a long process. We have the best rocket engineer team in history, and next year, I believe we'll finally be able to achieve full reusability of the rocket.